The Nature and Classification of Unlabelled Neurons in the Use of Kohonen's Self-Organizing Map for Supervised Classification

Kohonen’s Self-Organizing Map is a neural network procedure in which a layer of neurons is initialized with random weights, and subsequently organized by inspection of the data to be classified. The organization procedure uses progressive adjustment of weights based on data characteristics and lateral interaction such that neurons with similar weights will tend to spatially cluster in the neuron layer. When the SOM is associated with a supervised classification, a majority voting technique is usually used to associate these neurons with training data classes. This technique, however, cannot guarantee that every neuron in the output layer will be labeled, and thus causes unclassified pixels in the final map. This problem is similar to but fundamentally different from the problem of dead units that arises in unsupervised SOM classification (neurons which are never organized by the input data). Although dead unit problems have been widely discussed and some possible solutions proposed, few studies are associated with land cover classification. In this paper we specifically address the problem and nature of unlabelled neurons in the use of SOM for supervised classification. Through a case study it is shown that unlabelled neurons are associated with unknown image classes and, most particularly, mixed pixels. It is also shown that an auxiliary algorithm proposed here for assigning classes to unlabelled neurons performs with the same success as that experienced with Maximum Likelihood.